knowledge-rag 4.9.2__tar.gz → 4.10.0__tar.gz

{knowledge_rag-4.9.2 → knowledge_rag-4.10.0}/.gitignore

@@ -4,6 +4,10 @@ __pycache__/
 *.pyc
 data/
 
+# Package presets are versioned resources, not runtime storage.
+!/mcp_server/data/
+!/mcp_server/data/**
+
 # User config (personal settings — use presets/ as starting point)
 config.yaml
 

knowledge_rag-4.9.2/README.md → knowledge_rag-4.10.0/PKG-INFO

@@ -1,3 +1,52 @@
+Metadata-Version: 2.5
+Name: knowledge-rag
+Version: 4.10.0
+Summary: Local RAG System for Claude Code — Hybrid search + Cross-encoder Reranking + 13 MCP Tools + 20 Format Parsers. Zero external servers.
+Project-URL: Homepage, https://github.com/lyonzin/knowledge-rag
+Project-URL: Repository, https://github.com/lyonzin/knowledge-rag
+Project-URL: Issues, https://github.com/lyonzin/knowledge-rag/issues
+Project-URL: Changelog, https://github.com/lyonzin/knowledge-rag/releases
+Author-email: "Lyon." <lyonzin@users.noreply.github.com>
+License: MIT
+License-File: LICENSE
+Keywords: bm25,chromadb,claude-code,embeddings,fastembed,hybrid-search,knowledge-base,local-ai,mcp,rag,reranking,retrieval-augmented-generation,semantic-search
+Classifier: Development Status :: 4 - Beta
+Classifier: Intended Audience :: Developers
+Classifier: Intended Audience :: Science/Research
+Classifier: License :: OSI Approved :: MIT License
+Classifier: Operating System :: OS Independent
+Classifier: Programming Language :: Python :: 3.11
+Classifier: Programming Language :: Python :: 3.12
+Classifier: Programming Language :: Python :: 3.13
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
+Classifier: Topic :: Text Processing :: Indexing
+Requires-Python: >=3.11
+Requires-Dist: beautifulsoup4>=4.12.0
+Requires-Dist: chromadb>=1.4.0
+Requires-Dist: fastembed[reranking]>=0.4.0
+Requires-Dist: mcp<3.0.0,>=2.0.0
+Requires-Dist: numpy>=1.24.0
+Requires-Dist: openpyxl>=3.1.0
+Requires-Dist: pymupdf>=1.23.0
+Requires-Dist: python-docx>=1.0.0
+Requires-Dist: python-pptx>=1.0.0
+Requires-Dist: pyyaml>=6.0
+Requires-Dist: requests>=2.33.0
+Requires-Dist: watchdog>=4.0.0
+Provides-Extra: gpu
+Requires-Dist: nvidia-cublas-cu12<13,>=12.8; (sys_platform == 'win32' or sys_platform == 'linux') and extra == 'gpu'
+Requires-Dist: nvidia-cuda-runtime-cu12<13,>=12.8; (sys_platform == 'win32' or sys_platform == 'linux') and extra == 'gpu'
+Requires-Dist: nvidia-cudnn-cu12<10,>=9; (sys_platform == 'win32' or sys_platform == 'linux') and extra == 'gpu'
+Requires-Dist: nvidia-cufft-cu12; (sys_platform == 'win32' or sys_platform == 'linux') and extra == 'gpu'
+Requires-Dist: nvidia-curand-cu12; (sys_platform == 'win32' or sys_platform == 'linux') and extra == 'gpu'
+Requires-Dist: nvidia-cusolver-cu12; (sys_platform == 'win32' or sys_platform == 'linux') and extra == 'gpu'
+Requires-Dist: nvidia-cusparse-cu12; (sys_platform == 'win32' or sys_platform == 'linux') and extra == 'gpu'
+Requires-Dist: nvidia-nvjitlink-cu12; (sys_platform == 'win32' or sys_platform == 'linux') and extra == 'gpu'
+Requires-Dist: onnxruntime-gpu<1.27,>=1.21; (sys_platform == 'win32' or sys_platform == 'linux') and extra == 'gpu'
+Provides-Extra: server
+Requires-Dist: uvicorn>=0.20.0; extra == 'server'
+Description-Content-Type: text/markdown
+
 <!--
 knowledge-rag — Local RAG server for Claude Code, Cursor, Windsurf and other MCP clients.
 SEO keywords: local rag, mcp server, self-hosted rag, hybrid search, cross-encoder reranking, claude code rag, cursor rag, python rag, chromadb rag, fastembed rag, air-gapped rag, gdpr rag, retrieval augmented generation, model context protocol server, zero cloud rag.
@@ -80,14 +129,14 @@ pip install knowledge-rag   →   restart Claude Code   →   search_knowledge("
 
 Most RAG frameworks fall into one of three traps: (1) they require you to ship your data to a cloud API, (2) they hand you 300 building blocks and 0 opinionated defaults, or (3) they bundle RAG as a 5% feature of a much bigger platform you didn't ask for.
 
-**knowledge-rag does one thing well:** it is the MCP-native local RAG server that Claude Code, Cursor, Windsurf, VS Code, Cline, Gemini CLI and Zed can search out of the box — with enterprise plumbing (bearer auth, Prometheus metrics, rate limiting, health probes, structured JSON logging, zero-downtime reindex) that no other RAG-focused OSS ships built-in.
+**knowledge-rag is an MCP-native local RAG server** for clients such as Claude Code, Cursor, Windsurf, VS Code, Cline, Gemini CLI and Zed. It includes optional bearer authentication, Prometheus metrics, rate limiting, health probes, structured JSON logging and staging-based rebuilds. Configure the transport and access controls for the intended deployment.
 
 <table>
 <tr>
 <td width="50%" valign="top">
 
 ### 🔒 100% local, 0% cloud
-Your files never leave the machine. No vendor lock-in, no data-residency headache, no forced cloud dependency. **LGPD / GDPR / HIPAA compliant by architecture** — because there is nothing to comply about when nothing leaves.
+Embedding generation, indexing, and retrieval run locally. Model downloads and `add_from_url` use network access, and configured MCP clients receive retrieved content. Choose the corpus, client permissions, and network exposure to fit your deployment; this project does not certify regulatory compliance.
 
 </td>
 <td width="50%" valign="top">
@@ -101,7 +150,7 @@ Your files never leave the machine. No vendor lock-in, no data-residency headach
 <td width="50%" valign="top">
 
 ### 🛡️ Production-grade OSS
-7-pillar quality gate on every PR (**35+ automated checks**), 9-cell OS×Python CI matrix (Linux + Windows + macOS × 3.11/3.12/3.13), nightly chaos + 50K-iteration soak + mutation testing. **700+ tests. 0 known regressions.**
+7-pillar quality workflow, a 9-cell OS×Python CI matrix (Linux + Windows + macOS × 3.11/3.12/3.13), and nightly chaos, bounded-memory soak and mutation tests. Inspect the checks on the revision being deployed; CI coverage is not a guarantee of zero defects.
 
 </td>
 <td width="50%" valign="top">
@@ -117,31 +166,21 @@ No token bills. No SaaS tier. No paid features hidden behind a wall. **MIT licen
 
 ## 📊 How knowledge-rag compares to other RAG frameworks
 
-We audited **16 popular RAG frameworks and platforms** (LlamaIndex, LangChain, ChromaDB, Weaviate, Qdrant, RAGFlow, LightRAG, DSPy, GraphRAG, Haystack, RAG-Anything, kotaemon, txtai, llmware, Dify, open-webui, FastGPT) so you can pick honestly.
-
-**Legend:** ✅ built-in · 🟡 plugin / paid tier / partial · ❌ not available · ⚠️ license or default concern
-
-| Dimension | 🎯 **knowledge-rag** | LlamaIndex | LangChain | Haystack | RAGFlow | txtai | open-webui | Dify | Qdrant |
-|---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
-| **100% local, zero cloud** | ✅ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ✅ | 🟡 | 🟡 |
-| **MCP native (Claude/Cursor)** | ✅ 13 tools | 🟡 pkg | 🟡 adapter | 🟡 wrapper | 🟡 add-on | ✅ | ✅ consumer | ✅ | ❌ |
-| **Hybrid BM25 + semantic** | ✅ 128× faster | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ |
-| **Cross-encoder rerank** | ✅ builtin | ❌ | 🟡 | ✅ | ✅ fused | ❌ | ✅ | 🟡 | 🟡 |
-| **Bearer auth builtin** | ✅ | ❌ | ❌ | ❌ core | ❌ | 🟡 | ✅ RBAC | ✅ OAuth2 | ✅ |
-| **Prometheus `/metrics`** | ✅ | ❌ | ❌ | ❌ core | ❌ | ❌ | ✅ OTel | ❌ | ✅ |
-| **Rate limiting** | ✅ sliding-window | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ |
-| **Health probes** (`/health`) | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | 🟡 | 🟡 | ✅ |
-| **Structured JSON logging** | ✅ opt-in | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ OTel | 🟡 | ✅ |
-| **Zero-downtime reindex** | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ |
-| **Async background reindex** | ✅ + polling | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | 🟡 |
-| **GPU CUDA optional** | ✅ 12 auto | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | 🟡 |
-| **File formats builtin** | ✅ **20** | 0 (LlamaParse=$) | 50+ plugins | ✅ **36+** | 8+ | ? | ? | ~10 | ❌ |
-| **Setup < 5 min POC** | ✅ pip 1-liner | ✅ | ✅ | ✅ | ❌ 16GB RAM | ✅ | ✅ docker | ✅ docker | ✅ |
-| **Nightly chaos + soak + mutation** | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-| **License** | ✅ MIT | MIT | MIT | Apache-2.0 | Apache-2.0 | Apache-2.0 | ⚠️ preserving | ⚠️ restrictive | Apache-2.0 |
-
-> **The 5 dimensions where knowledge-rag is unique:** health probes + JSON logging + Prometheus + rate limit + bearer auth **simultaneously built-in on an OSS RAG-focused MCP server**. Zero-downtime reindex + async background reindex + nightly chaos/soak/mutation are documented on nobody else's README.
-
+Compare the features needed for your deployment rather than treating framework
+checklists as a benchmark. Knowledge-rag combines MCP tools, local embeddings,
+ChromaDB persistence, BM25, optional FTS5 and optional reranking in one process.
+
+| Decision | What to verify |
+| --- | --- |
+| Local or remote processing | Where models run, which client receives content, and which features need network access |
+| Retrieval quality | Expected-source queries on your corpus, including its languages and identifiers |
+| Performance | Cold initialization, indexing, uncached queries, memory peaks and concurrent use on comparable hardware |
+| Operations | Authentication, metrics exposure, backups, recovery and one-writer ownership |
+| Extensibility | Supported parsers, model profiles and the public MCP contract |
+
+The [audit report](docs/audit-2026-10-03.md) records a reproducible local workload
+and its limitations. Results against an older implementation of this project are
+not speed comparisons against other RAG frameworks.
 ---
 
 ## 🚀 Quick Start (3 minutes, from zero to your first query)
@@ -191,6 +230,30 @@ knowledge-rag --transport sse
 - Prometheus scrape: `http://your-host:9179/metrics`
 - MCP dispatcher: authenticated via `Authorization: Bearer your-secret-token`
 
+### Multi-client setup (Shared Server via Streamable HTTP)
+
+`stdio` is ideal when one MCP client owns one `knowledge-rag` process. If several clients (e.g. Claude, Cursor, LM Studio, or multiple terminal windows) should share one index and one running process, run `knowledge-rag` once with Streamable HTTP and point each HTTP-capable client at the same endpoint (prompted by [#217](https://github.com/lyonzin/knowledge-rag/pull/217) and issue [#223](https://github.com/lyonzin/knowledge-rag/issues/223)):
+
+```bash
+knowledge-rag --transport streamable-http
+# MCP endpoint: http://127.0.0.1:8179/mcp
+```
+
+- **Local MCP clients (Cursor, LM Studio, etc.):** Point their remote-server URL directly at `http://127.0.0.1:8179/mcp`.
+- **Claude Custom Connectors (Remote MCP):**
+  - **Network requirement:** Custom connector requests originate from Anthropic's cloud infrastructure across web, desktop, and mobile, so they cannot reach local `127.0.0.1`. You must expose your server via a secure, publicly accessible HTTPS endpoint (e.g. Cloudflare Tunnel, ngrok, or reverse proxy) at `https://your-public-domain/mcp`. Configure the proxy to send an upstream `Host` accepted by the HTTP app for `server.host` (for the default loopback configuration, use `127.0.0.1:8179`); forwarding the public hostname unchanged can produce `421 Invalid Host header`. Keep bearer authentication and HTTPS enabled. Local `claude_desktop_config.json` uses `command`/`args` for local stdio processes and does not support remote HTTP endpoints.
+  - **Setup & Auth:** In Claude (Pro/Max plans: **Customize > Connectors > + Add > Add custom connector**; Team/Enterprise owners: **Organization settings > Connectors**), enter your public endpoint. In **Request headers**, set header name `Authorization` with value `Bearer <your-token>` matching `server.auth.bearer_token` in `config.yaml` (see [Anthropic Custom Connectors documentation](https://support.claude.com/en/articles/11175166-get-started-with-custom-connectors-using-remote-mcp)).
+
+| Transport | Clients | Shared index / process | RAM overhead | Process management | Startup & query latency | Best fit |
+|---|:---:|:---:|---|---|---|---|
+| `stdio` | 1 | No | 1 runtime per client; models load lazily in each process when queried | Client manages its own process; no separate server | Spawns per client; pays preflight and first-query model load | Simplest local setup (single client) |
+| `sse` | Many | Yes | Shared server process; single shared model load across clients | Operator manages one background server | Reuses warm server; first-query model load still applies | Legacy HTTP clients |
+| `streamable-http` | Many | Yes | Shared server process; single shared model load across clients | Operator manages one background server | Reuses warm server; first-query model load still applies | **Preferred** shared-server setup |
+
+When an HTTP transport is selected, `knowledge-rag` automatically enables its single-instance lock (`KNOWLEDGE_RAG_SINGLE_INSTANCE=1`) to prevent multiple servers from binding the same data directory or port.
+
+> **Security note:** The default host is `127.0.0.1`. If you bind to `0.0.0.0` or deploy across a network, configure `server.auth.bearer_token` in `config.yaml` and terminate TLS (HTTPS) via a reverse proxy (e.g. Caddy, Nginx) to protect bearer tokens in transit; otherwise the server warns that the HTTP endpoint is unauthenticated.
+
 ### Path 3 — Docker (models pre-downloaded, air-gapped ready)
 
 ```bash
@@ -273,8 +336,8 @@ Every RAG framework claims "production-ready." Here is what knowledge-rag ships
 
 - **Bearer token auth** on SSE / HTTP transports — constant-time comparison (`hmac.compare_digest`), RFC 6750 challenge, 401 fenced with `WWW-Authenticate` header
 - **Path traversal + symlink escape defenses** — `validate_path_within` guarding 6 CRUD tools (CWE-22, CWE-59)
-- **Prompt injection 3-layer defense** — sentinel neutralization + provenance fence + `external_source` flag (OWASP LLM01:2025)
-- **OpenSSF Best Practices badge** verified · **CodeQL** weekly scan · **Bandit + Semgrep + Gitleaks + pip-audit** on every PR
+- **External-content handling** — known-token neutralization and a stored provenance fence, with parser metadata propagation. These are limited mitigations; clients must still treat retrieved content as untrusted data. See [SECURITY.md](SECURITY.md).
+- **OpenSSF Best Practices** project [#13864](https://www.bestpractices.dev/en/projects/13864) · **CodeQL** weekly scan · **Bandit + Semgrep + Gitleaks + pip-audit** workflows. Review actual check results and exceptions for the release you deploy.
 - **PyPI Trusted Publishing** via OIDC (zero long-lived tokens in CI)
 
 </td>
@@ -296,12 +359,12 @@ Every RAG framework claims "production-ready." Here is what knowledge-rag ships
 ### Scale & performance
 
 - **SSE / streamable-http transport** — 1 server serves N MCP clients, ChromaDB WAL mode enabled automatically, shared embedding model + query cache
-- **BM25 inverted-index** — **128× faster** than linear scan (custom implementation, replaces `rank-bm25`)
-- **FTS5 SQLite fast-path** (opt-in, ADR-002/003/006/008) — <10ms cold, <2ms hot on lexical queries
-- **Cross-encoder reranking** — Xenova/ms-marco-MiniLM-L-6-v2, +1.88pp Recall@10 (p<0.001)
-- **GPU CUDA 12** with auto DLL discovery + graceful CPU fallback
-- **Query cache** — LRU + 5-min TTL, cuts p95 latency ~40%
-- **Zero-downtime reindex** — staging populate + validation + atomic swap + durable metadata rollback
+- **BM25 inverted index** — keyword candidate retrieval without scoring every document for every query
+- **FTS5 SQLite fast path** — opt-in lexical retrieval; measure relevance and latency on your identifiers and corpus
+- **Cross-encoder reranking** — Xenova/ms-marco-MiniLM-L-6-v2; optional model scoring of a candidate pool
+- **NVIDIA CUDA embeddings** and experimental Windows AMD DirectML opt-in, with actual-session checks and CPU fallback; [setup and compatibility limits](docs/gpu-setup.md)
+- **Query cache** — LRU + 5-minute TTL, with mutation invalidation
+- **Staging rebuild** — populate and validate a replacement while the current collection remains available; handled publication failures roll back
 - **Async background reindex** with `get_reindex_status()` polling
 
 </td>
@@ -310,11 +373,11 @@ Every RAG framework claims "production-ready." Here is what knowledge-rag ships
 ### Reliability
 
 - **Nightly chaos injection** — HuggingFace Hub offline · ONNX zero-byte replay · watchdog crash recovery (3 scenarios in `tests/chaos/`)
-- **50 000-iteration soak test** — proves no memory leak after 1h of continuous queries (`KNOWLEDGE_RAG_SOAK_ITERATIONS=50000`)
+- **50,000-iteration soak test** — checks a fixed memory-growth budget with stub embeddings (`KNOWLEDGE_RAG_SOAK_ITERATIONS=50000`); this is not a one-hour real-model endurance test
 - **Mutation testing** (mutmut) on `instance_lock` + `preflight` — catches tests that are too weak
 - **Determinism check** — full test suite × 3, catches flakes
-- **Backwards-compat frozen** — 13 MCP tool parameter names guarded by `tests/test_backwards_compat.py` + legacy YAML fixtures (v3.6.0 / v3.7.0) still parse
-- **API surface AST diff** — `check_api_surface.py` blocks any breaking change at PR time
+- **Backwards compatibility** — 13 MCP tools, signature regression tests and legacy YAML fixtures; optional additions preserve existing calls
+- **API surface AST diff** — checks structural signature changes alongside behavioral tests
 - **9-cell CI matrix** — Linux + Windows + macOS × 3.11 + 3.12 + 3.13
 
 </td>
@@ -329,7 +392,7 @@ Every RAG framework claims "production-ready." Here is what knowledge-rag ships
 
 **Preset:** [`cybersecurity.yaml`](presets/cybersecurity.yaml) · 8 categories · 200+ routing keywords · 69 query expansions
 
-Ingest MITRE ATT&CK, threat reports, exploit writeups, incident reports. Search from Claude Code with `search_knowledge("privilege escalation windows")` and get instant recall across your entire corpus. Air-gapped — nothing leaves the laptop.
+Ingest MITRE ATT&CK, threat reports, exploit writeups and incident reports. Search from an MCP client with `search_knowledge("privilege escalation windows")`. Offline operation requires cached models; the chosen client receives retrieved content and determines what happens to it afterwards.
 
 ### Development Teams — Design Docs, Runbooks, Code
 
@@ -388,8 +451,8 @@ flowchart TB
 
     subgraph SEARCH["Retrieval Pipeline"]
         R[Query Router<br/>lexical vs semantic]
-        F[FTS5 Fast-Path<br/>opt-in · lt 10ms]
-        BM[BM25 Inverted Index<br/>128x faster than baseline]
+        F[FTS5 Fast-Path<br/>opt-in lexical retrieval]
+        BM[BM25 Inverted Index<br/>keyword candidates]
         SE[Semantic Search<br/>FastEmbed ONNX lazy-loaded]
         RRF[Reciprocal Rank Fusion]
         CE[Cross-Encoder Rerank<br/>MiniLM-L-6-v2]
@@ -405,11 +468,11 @@ flowchart TB
     subgraph INGEST["Document Ingestion"]
         FS[documents/ folder]
         WD[Watchdog<br/>10s debounce]
-        PA[35 Parsers<br/>MD · PDF · DOCX · code · IaC · IPYNB]
+        PA[Format parsers<br/>MD · PDF · DOCX · code · IaC · IPYNB]
         CK[Chunker<br/>markdown-aware · code-aware]
         EM[FastEmbed ONNX<br/>384D bge-small-en-v1.5]
         DD[SHA256 Dedup]
-        SW[Zero-downtime Staging Swap<br/>rollback on validation fail]
+        SW[Staging Rebuild<br/>rollback on handled publication failure]
     end
 
     subgraph OBS["Enterprise Observability (opt-in)"]
@@ -448,7 +511,7 @@ flowchart TB
 
     FS --> WD --> PA
     PA --> CK --> EM --> DD --> CH
-    SW -.->|atomic swap| CH
+    SW -.->|validated publication| CH
     SW -.-> FT
     CH -.-> MD
 
@@ -480,26 +543,26 @@ flowchart TB
 
 1. **Any MCP client** — Claude Code, Cursor, Windsurf, and 5 others — connects via the transport of your choice (stdio for personal use, SSE/streamable-http for teams).
 2. **HTTP mode chains 3 ASGI middlewares in order**: health probes first (always answered), then bearer auth (fenced with `WWW-Authenticate`), then rate limiter (sliding window).
-3. **All 13 MCP tools** are decorated with `@rate_limited` + `@instrument` — Prometheus counts every call, rate limiter enforces RPM+burst, both zero-cost when disabled.
+3. **All 13 MCP tools** are decorated with `@rate_limited` + `@instrument` — metrics record calls and the rate limiter enforces RPM+burst when enabled. Disabled features skip that work, while the lightweight wrapper calls remain.
 4. **`search_knowledge` checks the query cache first**; cache miss routes through the Query Router (regex classifier) to either the FTS5 fast-path (lexical) or the hybrid pipeline (BM25 + semantic + RRF + cross-encoder rerank).
 5. **Storage is 100% local**: ChromaDB (WAL mode) for vectors + metadata, SQLite FTS5 (WAL + busy-timeout) for lexical fast-path, `index_metadata.json` for durable state.
-6. **Document ingestion runs continuously**: watchdog observes `documents/`, 35 parsers handle each format, chunker respects language boundaries, FastEmbed ONNX generates embeddings, SHA256 deduplicates, and a staging swap performs zero-downtime rebuilds with rollback-on-failure.
+6. **Document ingestion**: when enabled, watchdog observes `documents/`; format parsers create chunks, FastEmbed generates embeddings, and source metadata supports incremental updates. Full rebuilds use staging. See the operations guide for publication and recovery limits.
 7. **Enterprise observability** (opt-in) — Prometheus `/metrics`, structured JSON logs, `/health` payload — attaches to the same instrumentation points, no code changes required.
 8. **`config.yaml` (with 5 domain presets)** controls every subsystem — no environment variable spaghetti, no hardcoded paths.
 
-**Complete architecture — 4 detailed Mermaid diagrams (System Overview · Query Flow · Document Ingestion · hybrid_alpha effect):** [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md)
+**Component ownership, query/ingestion diagrams, memory boundaries and hybrid weights:** [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md)
 
 ---
 
 ## 📄 35 File Formats — parsed natively, no plugins needed
 
-Every parser is **chunk-aware** — Markdown splits at `##` headers, code splits at function/class boundaries, notebooks skip base64 outputs, PDFs use PyMuPDF, spreadsheets extract sheet-by-sheet. **33 formats are enabled by default**; the 2 MetaTrader formats are opt-in (add to `documents.supported_formats` in `config.yaml`).
+Parsers use format-specific extraction: Markdown sections, code boundaries, notebook cells without base64 outputs, PDF text and spreadsheet rows. Several extensions share one parser. The runtime fallback list contains 33 formats; the documented template and presets can select a smaller list. Inspect your effective `documents.supported_formats`. The two MetaTrader extensions are opt-in.
 
 | # | Format | Extension | Parser | Default | Notes |
 |---|--------|-----------|--------|:-------:|-------|
 | 1 | Markdown | `.md` | Section-aware (splits at `##`) | Yes | Headers preserved as chunk boundaries |
 | 2 | Plain Text | `.txt` | Fixed-size chunking | Yes | 1000 chars + 200 overlap |
-| 3 | PDF | `.pdf` | PyMuPDF extraction | Yes | Text-based PDFs only (no OCR) |
+| 3 | PDF | `.pdf` | PyMuPDF extraction or explicit page-marked UTF-8 text | Yes | No OCR; extracted-text recognition is limited to 4 MiB |
 | 4 | Word | `.docx` | python-docx | Yes | Headings preserved as markdown |
 | 5 | Excel | `.xlsx` | openpyxl | Yes | Sheet-by-sheet extraction |
 | 6 | PowerPoint | `.pptx` | python-pptx | Yes | Slide-by-slide extraction |
@@ -533,6 +596,16 @@ Every parser is **chunk-aware** — Markdown splits at `##` headers, code splits
 | 34 | MQL4 Source | `.mq4` | Code parser | **No** | MetaTrader — opt-in via `documents.supported_formats` |
 | 35 | MQL4 Header | `.mqh` | Code parser | **No** | MetaTrader — opt-in via `documents.supported_formats` |
 
+Files with a `.pdf` extension can also contain an explicit text extraction: the
+first line must be `[Page N]` with a positive page number, followed by nonempty
+UTF-8 text. An optional UTF-8 BOM and tab/CR/LF whitespace are accepted. This
+recognition is limited to 4 MiB, rejects other C0/C1 controls and DEL, and never
+reinterprets a file containing PDF magic in the first 1 KiB as text. Accepted
+extractions retain their source path and are marked `content_format=extracted_text`
+and `page_count_source=markers`; the marker count is not proof of the original
+PDF's total page count. Binary PDFs keep the native parser and its error
+diagnostics. Invalid or oversized textual lookalikes remain errors.
+
 > **Enable an opt-in format** — add the extension to `documents.supported_formats` in your `config.yaml`:
 > ```yaml
 > documents:
@@ -588,7 +661,7 @@ Native MCP
 
 **Complete client configuration guide with JSON schemas per client:** [docs/INSTALLATION.md#use-with-other-mcp-clients →](docs/INSTALLATION.md#use-with-other-mcp-clients)
 
-> ⚠️ **Operating model — read once.** Your MCP client (Claude Desktop, Code, Cursor, etc.) **spawns and owns the knowledge-rag server process**. Do **NOT** also run `knowledge-rag` in a terminal while your MCP client is open — two writer processes against the same `data_dir` corrupt the ChromaDB HNSW segment on Windows (see [issue #216](https://github.com/lyonzin/knowledge-rag/issues/216)). Recommended baseline: add `"KNOWLEDGE_RAG_SINGLE_INSTANCE": "1"` to the `env` block of your MCP config as a safety net. Full explanation + how to run CLI commands safely: [docs/single-instance.md](docs/single-instance.md).
+> ⚠️ **Operating model — read once.** In `stdio` mode, your MCP client (Claude Desktop, Code, Cursor, etc.) **spawns and owns the knowledge-rag server process**. Do **NOT** also run a second `knowledge-rag` writer in a terminal while that stdio client is open — two writer processes against the same `data_dir` corrupt the ChromaDB HNSW segment on Windows (see [issue #216](https://github.com/lyonzin/knowledge-rag/issues/216)). For HTTP transports, manually running one server and connecting HTTP clients to it is the supported shared-server workflow; do not mix it with another writer using the same `data_dir`. Recommended baseline for stdio clients: add `"KNOWLEDGE_RAG_SINGLE_INSTANCE": "1"` to the `env` block of your MCP config as a safety net. Full explanation + how to run CLI commands safely: [docs/single-instance.md](docs/single-instance.md).
 
 ---
 
@@ -604,7 +677,8 @@ paths:
 models:
   embedding:
     profile: "compact"                  # "compact" | "quality" | "multilingual" | "custom"
-    gpu: "auto"                         # "auto" | "true" | "false"
+    gpu: "auto"                         # "auto" | "true" | "false" | "directml"
+    device_id: null                      # DirectML: explicit DXGI adapter index
   reranker:
     enabled: true                       # cross-encoder rerank
 
@@ -642,12 +716,12 @@ knowledge-rag is designed for teams that cannot let their documents leave the pe
 | **Supply chain security** | PyPI Trusted Publishing via OIDC (no long-lived tokens) |
 | **Vulnerability disclosure** | Private security advisory via [SECURITY.md](SECURITY.md) |
 | **Signed release attestations** | GitHub release attestations on every published version |
-| **Reproducible builds** | Locked `requirements.txt` with pinned versions |
+| **Dependency management** | Version ranges in `requirements.txt` / `pyproject.toml`; resolve and record a tested environment for reproducibility |
 | **Authenticated access** | Bearer token middleware on SSE / HTTP transports (constant-time compare, RFC 6750) |
 | **Rate limiting** | Sliding-window per-client RPM + burst (opt-in, zero-cost when disabled) |
 | **Audit-ready logging** | Opt-in structured JSON logs → ship to your SIEM |
 | **Path traversal defenses** | CWE-22 / CWE-59 guards on 6 CRUD tools |
-| **Prompt injection defense** | 3-layer sanitization on `add_from_url` (OWASP LLM01:2025) |
+| **External-content mitigation** | Known-token neutralization and provenance fencing on `add_from_url`; not a general prompt-injection prevention guarantee |
 
 **OpenSSF Best Practices** badge: passing · project ID [#13864](https://bestpractices.coreinfrastructure.org/projects/13864)
 
@@ -655,16 +729,14 @@ knowledge-rag is designed for teams that cannot let their documents leave the pe
 
 ## 📈 Numbers that matter
 
-- **26 000+** total downloads on PyPI · **250+** GitHub stars · **70+** enterprise teams (private + community)
-- **700+ tests** collected · **1.33:1** test-to-code ratio · **codecov trend gate** ±0.5pp
-- **35+ status checks** on every PR (9-cell OS×Python matrix · 7 quality pillars)
-- **35 file formats** parsed natively · **13 MCP tools** frozen · **5 domain presets** (cyber · dev · research · multilingual · general)
-- **BM25 128× faster** than baseline · **cross-encoder +1.88pp** Recall@10 (p<0.001) · **cache −40%** p95 latency
-- **1 800+ files / 39 K chunks indexed in < 3 min** on a modern laptop (typical developer corpus)
-- **Verified in production** on 5 889-doc / 75 016-chunk corpora
-
-**Public benchmark dashboard:** https://lyonzin.github.io/knowledge-rag/
+- **13 MCP tools**, configurable format selection and five domain presets.
+- **Nine OS/Python CI combinations** configured for Linux, macOS and Windows with Python 3.11–3.13.
+- **Measured CPU/CUDA indexing and retrieval, AMD validation and memory regressions:** [audit results and reproduction](docs/audit-2026-10-03.md).
+- **Historical implementation benchmarks:** [v4.7.1 baseline](docs/perf-baseline-v4.7.1.md). Historical ratios apply to their recorded workloads, not every corpus or competing product.
 
+The [public benchmark dashboard](https://lyonzin.github.io/knowledge-rag/) tracks
+CI workloads. Mocked storage/model workloads and real-model measurements answer
+different questions; inspect the methodology and units before comparing them.
 ---
 
 ## 📚 Documentation
@@ -674,15 +746,16 @@ knowledge-rag is designed for teams that cannot let their documents leave the pe
 | [**Installation guide**](docs/INSTALLATION.md) | 5 install methods · 8 MCP client integrations · GPU setup |
 | [**API reference**](docs/API.md) | Complete reference for all 13 MCP tools |
 | [**Configuration reference**](docs/CONFIGURATION.md) | Every `config.yaml` field · presets · tuning |
-| [**Architecture**](docs/ARCHITECTURE.md) | 4 Mermaid diagrams: System Overview · Query Flow · Ingestion · hybrid_alpha |
+| [**Architecture**](docs/ARCHITECTURE.md) | Components · query/ingestion flow · memory/concurrency boundaries · hybrid weights |
 | [**Troubleshooting**](docs/TROUBLESHOOTING.md) | 11 common issues + solutions |
 | [**FTS5 fast-path guide**](docs/features/fts5_fast_path.md) | Opt-in lexical fast-path — when and how |
 | [**Reindex operations**](docs/reindex-operations.md) | Zero-downtime rebuild · resume · checkpoint |
-| [**GPU setup**](docs/gpu-setup.md) | CUDA 12 installation + troubleshooting |
+| [**GPU setup**](docs/gpu-setup.md) | CUDA runtime compatibility · CPU fallback · AMD/macOS GPU limitations |
 | [**Migration to v4.8.0**](docs/migration-v4.8.0.md) | Embedding profile · multilingual · zero-downtime |
 | [**Security policy**](SECURITY.md) | Threat model · disclosure channel |
 | [**Contributing**](CONTRIBUTING.md) | Development · testing · PR process |
 | [**Changelog**](CHANGELOG.md) | All release notes since v1.0.0 |
+| [**Memory/GPU audit**](docs/audit-2026-10-03.md) | Reproducible workloads · measured results · validation limits · open PR/issue review |
 
 ---